Feature Selection Using Enhanced Particle Swarm Optimisation for Classification Models.

Area of Science:

• Computational Intelligence
• Machine Learning
• Optimization Algorithms

Background:

• Particle Swarm Optimisation (PSO) is a metaheuristic optimization algorithm widely used for various tasks.
• The original PSO model suffers from premature convergence and insufficient exploitation of near-optimal solutions.
• Effective feature selection is crucial for improving model performance and reducing computational complexity.

Purpose of the Study:

• To propose two novel variants of Particle Swarm Optimisation (PSO) tailored for feature selection.
• To address the limitations of premature convergence and weak exploitation inherent in the standard PSO algorithm.
• To enhance the efficiency and effectiveness of discriminative feature selection.

Main Methods:

• Development of a first PSO variant incorporating modified PSO operations, spiral search local exploitation, Gaussian distribution-based swarm leader enhancement, and mirroring/mutation operations.
• Enhancement of the first variant into a second PSO model featuring an adaptive exemplar breeding mechanism, nonlinear function-oriented search coefficients, and advanced swarm leader/worst solution strategies.
• Comparative analysis against 15 classical and advanced search methods.

Main Results:

• The proposed PSO variants demonstrated statistically superior performance in discriminative feature selection.
• Evaluated across a diverse set of 13 datasets, the enhanced models consistently outperformed existing methods.
• The novel strategies effectively mitigated premature convergence and improved solution exploitation.

Conclusions:

• The developed PSO variants offer significant improvements over the original PSO for feature selection tasks.
• These enhanced algorithms provide a robust and effective approach for discriminative feature selection.
• The proposed methods represent a valuable contribution to the field of optimization and machine learning.